Voices
Create a voice for speech generation by describing it or cloning it from recordings, share it with your workspace, and use it in canvas, chat and MCP.
Settings → Workspace → Voices holds the voices used for speech generation. It lists Workspace voices (yours and those teammates shared, badged Described or Cloned, Shared or Private) and Endless voices, which every workspace can use and no one can edit.
Create a voice
Press "Create voice". You need permission to create content, so a Viewer or Billing manager can use shared voices but not create them.
| Method | How it works | Voice slot |
|---|---|---|
| Create by prompting | Designed from a written description, in under a minute. | No |
| Create from audio references | Cloned from one person's recordings, in about a minute. | Yes |
By prompting
Describe it
In "What it sounds like", describe age, accent, tone and pace in at least 20 characters. English works best.
Name and share
Enter a Name. Turn on "Share with the workspace" so teammates can use it; otherwise only you see it.
Press "Create voice"
If the description doesn't work, the error shows while the form is still open.
A described voice costs a few extra credits each time it speaks, because its sample is generated again.
From recordings
- Add Recordings of one person speaking, without music or background noise. 1 to 2 minutes in total works best. Only recordings you uploaded or generated yourself can be used.
- Pick the Provider. The voice works with that provider's speech models.
- Enter a Name and choose whether to share it.
- Tick the consent box confirming the person agreed to have their voice cloned.
A workspace can hold three cloned voices, counting every cloned voice whoever made it and whether it's shared. At the limit, delete one or create a described voice instead.
Share or delete a voice
Only the voice's creator sees its menu: "Share with the workspace" or "Make private", and "Delete voice". Deleting removes it for everyone. Audio it made and the recordings it was cloned from stay. If you move a training recording to the trash, the voice stays, but it can't be retrained from that recording once the trash is emptied.
Use a voice
- Canvas: the speech model picker lists Endless voices and the workspace voices you can see. "Create new voice" opens the same form.
- Chat: ask for a voice by name. The chat can also create a voice, and clones only after you confirm consent.
- MCP: connected agents can list, use and create voices. Cloning still needs consent.
A voice is available only where its speech model is: if Model availability turns the model off for you or it needs a higher plan, its voices are unavailable too.